The project WhyBehindScenes aims at developing new methods for automatically understanding the storyline in videos, and in particular the why behind the scenes in edited videos (films and TV shows). This will be investigated in two directions: first, by automatically understanding the storyline by focusing on the timeline, parts that are crucial for a plot, and correlation of scenes; second, by identifying the filmmaker’s decisions and integrating them in the video analysis by extracting the film’s signature (the style of the direction, the emotion a scene is conveying) and by creating novel sequences with specific directorial style. WhyBehindScenes targets to build systems capable of converting videos to books, by including not only the audio-visual information present in scenes but also the motivation, intention or even emotion behind events.

The FiVASt project focuses on the approximation of stochastic partial differential equations (SPDEs) using finite-volume methods. The aim is to carry out a complete numerical analysis of these problems (existence and uniqueness of an approximate solution, convergence of this solution towards the solution of the initial problem, etc.) as well as numerical simulations using finite-volume methods having already proven their efficiency in the deterministic framework. This project will then have to couple a deterministic analysis of finite-volume methods with tools specific to stochastic calculus. Indeed, even if more and more fields show that stochastic or random effects are necessary for a good description of reality, unlike models based on stochastic differential equations, SPDEs are still poorly understood and many questions arise: restricted type of equation, few models and numerical approaches. In this context, the study of SPDEs is a currently booming subject due to the interest that they bring to the enrichment of existing models. The development of methodological tools from both a theoretical and numerical point of view seems to be a priority objective to effectively understand the use of models including stochastic aspects. In this sense, we propose a research project at the interface between the analysis of partial differential equations, numerical analysis and stochastic calculus. The goal is then to study finite-volume approximations for stochastic problems, thereby providing a comprehensive framework for the rigorous study of larger and more complex problems in the future.

The lungs are the primary organs of the respiratory system in humans and many animals, responsible for molecular exchanges between external air and internal blood through mechanical ventilation. It has an extraordinary complex architecture, with the inherent fractal structure of the bronchial and blood vessel trees, as well as the hierarchical structure of the parenchyma. Lung biomechanics has been extensively studied by physiologists, experimentally as well as theoretically, from the air flow, blood flow and tissue stress points of view, laying the ground for our current fundamental understanding of the relationship between function and mechanical behavior. However, many questions remain, notably in the intricate coupling between the multiple constituents, between the many phenomena taking place at different spatial and temporal scales in health and disease. For example, even for healthy lungs, there is no quantitative model allowing to link tissue-level and organ-level experimental material responses. These fundamental questions represent real clinical challenges, as pulmonary diseases are an important health burden. Interstitial lung diseases, for instance, affect several million people globally. Idiopathic Pulmonary Fibrosis, notably, a progressive form of interstitial lung diseases where some alveolar septa get thicker and stiffer while others get completely damaged, remains poorly understood, poorly diagnosed, and poorly treated, with a current median survival rate inferior to 5 years. It has, however, been hypothesized that a mechanical vicious cycle is in place within the parenchyma of IPF patients, where fibrosis and damage induce large stresses, which in turns favor fibrosis. The general goal of this project is twofold: (i) scientifically, to better understand pulmonary (solid) mechanics, from the alveolar scale to the organ in health and (IPF) disease; (ii) clinically, to improve diagnosis and prognosis of (IPF) patients through personalized computational modeling. More precisely, I propose to develop a many-scale model of the parenchymal biomechanics, at all relevant spatial scales from the alveolus to the organ, and at the temporal scales of the breathing cycle and fibrosis process. Different representations at successive spatial scales will be linked by a computational nonlinear homogenization strategy with a priori model reduction based on a neural network. The model will integrate the rather unique experimental data produced by Drs. Bel-Brunon and Trunfio-Sfarghiu from LaMCoS (INSA-Lyon), i.e., 30 microtomography images at alveolar scale, plus 10 inflation tests of lobules: microstructures will be extracted from the images and systematically analyzed, and model parameters will be estimated from the mechanical tests. The model will also integrate clinical-radiological data provided by Profs. Nunes and Brillet from Avicenne APHP Hospital, i.e., standard pulmonary function tests and thoracic computed tomography imaging on 10 IPF patients plus 5 normal lung controls: a pipeline to estimate observable model parameters from clinical data will be set up, and generic values will be defined for the remaining parameters. The model and estimation procedure will represent augmented diagnosis and prognosis tools for the clinicians. The project will be coordinated by Dr. Genet, who is currently an Assistant Professor in the Mechanics Department of École Polytechnique with research posting within the M?DISIM team, which belongs to both INRIA and the Solid Mechanics Laboratory of École Polytechnique/CNRS. Throughout the project he will be assisted by Drs. Chapelle and Moireau at INRIA/École Polytechnique, and maintain strong scientific collaborations with the LaMCoS at INSA-Lyon and Télécom-SudParis, as well as strong clinical collaborations with the Avicenne APHP Hospital and Hypoxia & Lung Laboratory of Paris XIII University/INSERM.

This is a project for higher education student and staff mobility between Programme Countries and Partner Countries. Please consult the website of the organisation to obtain additional details.

We propose to lay the theoretical foundations and design efficient computational methods for analyzing, quantifying and exploring relations and variability in structured data sets, such as collections of geometric shapes, point clouds, and large networks or graphs, among others. Unlike existing methods that are tied and often limited to the underlying data representation, our goal is to design a unified framework in which variability can be processed in a way that is largely agnostic to the underlying data type. In particular, we propose to depart from the standard representations of objects as collections of primitives, such as points or triangles, and instead to treat them as functional spaces that can be easily manipulated and analyzed. Since real-valued functions can be defined on a wide variety of data representations and as they enjoy a rich algebraic structure, such an approach can provide a completely novel unified framework for representing and processing different types of data. Key to our study will be the exploration of relations and variability between objects, which can be expressed as operators acting on functions and thus treated and analyzed as objects in their own right using the vast number of tools from functional analysis in theory and numerical linear algebra in practice. Such a unified computational framework of variability will enable entirely novel applications including accurate shape matching, efficiently tracking and highlighting most relevant changes in evolving systems, such as dynamic graphs, and analysis of shape collections. Thus, it will permit not only to compare or cluster objects, but also to reveal where and how they are different and what makes instances unique, which can be especially useful in medical imaging applications. Ultimately, we expect our study to create to a new rigorous, unified paradigm for computational variability, providing a common language and sets of tools applicable across diverse underlying domains.